I have noticed an alarmingly high incidence of TCP retransmissions from NIC on
FTP transfers over medium-delay, low-loss paths. On a typical transfer from
its ARPANET address via an ARPANET gateway and 4800-bps line for instance, an
average of 2.5 packet transissions occured for every packet stored. During this
transfer no packets were recorded as lost in either the gateway or destination
host - all spurious packets were duplicates. In transfers to an Ethernet host
connected via the same path an average of 2.0 packet transmissions occured for
every packet stored.
I supsect something is wrong in the retransmission-timeout algorithm of the NIC
TCP client, since other TOPS-20 hosts do not exhibit anything like this behavior;
however, other TOPS-20s seem on the hairy edge of eagerness, since ISI hosts,'f
for instance, seem to generate rather more duplicates that, say, Unix hosts over
the same paths. Fuzzball hosts (ahem) generate next to none over most paths. It
implementors read RFC-889.
It would be useful to learn if others can confirm or refute these observations,
since NIC is fast becoming a major fountain of data.
This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:36:06 GMT